Composite Structures 131 (2015) 905–916

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Probabilistic micromechanical analysis of composite material stiffness

properties for a wind turbine blade
Ghulam Mustafa ⇑, Afzal Suleman, Curran Crawford
University of Victoria, Department of Mechanical Engineering, P.O. Box 3055, Victoria, BC V8W 3P6, Canada

a r t i c l e i n f o a b s t r a c t

Article history: This work presents a coupled approach for stiffness property prediction of composite materials used in
Available online 4 July 2015 wind turbine blades using an advanced micromechanics and reliability-based methodologies. This
approach demonstrates how to map the uncertainties in the fiber and matrix properties onto the equiv-
Keywords: alent stiffness properties of composite laminates. Square and hexagonal unit cells were employed for the
Micromechanics estimation of the composite equivalent properties. The finite element formulation of the unit cells were
Unit cell performed in the ANSYS Multiphysics. The results from numerical experimentation conform well with
Reliability
the available test data and to the results from the Modified Rule of Mixture (MROM). A probabilistic anal-
Composite
Fiber
ysis using Monte Carlo Simulation with Latin Hypercube Sampling was used to assess the uncertainties in
Monte Carlo the equivalent properties according to the variability in the basic properties of the constituents.
Furthermore, a sensitivity analysis based on the Spearman Rank Order correlation coefficients was car-
ried out to highlight the influence of important properties of the constituents. As an illustration, the
above approach is applied to analyze a 5 MW wind turbine blade section under static loading. Results
demonstrate the possibility of the coupled approach at macro level (structure) from micro level (unit cell)
with the aim to design robust structures.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introductions manufacturing environment. On top of this, manufacturers usually

The size of wind turbines is increasing to capture more energy
and this trend will continue into the future [1]. Several
multi-MW prototype wind turbines exist for offshore applications
[2,3]. The power of a wind turbine scales as square of the rotor
diameter, while the mass of the blade (for similar conceptual
designs) scales as the cube of rotor diameter. Considering these
scaling laws one might predict that in the end material costs would
govern and avert further scaling. The cost of the blade would then
also scale as the cube of rotor diameter but advanced structural
concepts reduce its scaling exponent to 2.5 [4]. A
polymer-based composite material is a good choice for large struc-
tures such as wind turbine blades. The high strength-to-density
ratio, high stiffness-to-density ratio, good fracture toughness, fati-
gue performance and suitability for use in fast production of large
structures makes composites a good choice for their use in struc-
tural applications. The composite properties provided by the man-
ufacturer are generally the average properties in a particular
⇑ Corresponding author.
E-mail addresses: gmustafa@uvic.ca (G. Mustafa), suleman@uvic.ca
(A. Suleman), curranc@uvic.ca (C. Crawford).
don’t mention the number of tests that they performed to obtain
the average. This adds risk to structural design, especially for
large-scale composite layups such a blades. It is impossible to com-
pletely control variation in the composite properties. There are
many factors that contribute to that variation including:
batch-to-batch production, manufacturing conditions like temper-
ature, pressure, and humidity, curing time, labor skill, along with
the property variations in the basic building blocks of the ply
(the fiber and the matrix). These variations in the mechanical prop-
erties of composite materials are due primarily to variation in the
properties of the constituents - the fiber and matrix [5]. This vari-
ation in properties establishes a scatter in the response of a struc-
ture made up from this material, for example, the deflection of the
wind turbine blade. Therefore, it is necessary to consider the vari-
able nature of composite properties at the design stage. Current
wind turbine blade design is based on a deterministic approach
with a large factor of safety to ensure target static and fatigue lim-
its [6]. It is very expensive experimentally to obtain the design
allowables for a composite laminate in a deterministic approach
as significant test campaigns for each candidate layup are required.
The micromechanics (MM) based homogenization approach is
good alternative to characterize the stiffness properties of

http://dx.doi.org/10.1016/j.compstruct.2015.06.070
0263-8223/Ó 2015 Elsevier Ltd. All rights reserved.
906 G. Mustafa et al. / Composite Structures 131 (2015) 905–916
composite materials [7,8]. The Rule of Mixtures (ROM), also known
as the Simple Rule of Mixture (SROM), is one of the oldest and sim-
plest forms of micromechanics for calculating mechanical proper-
ties of unidirectional plies [7,9,10]. Halphin [11] proposed a
Modified Rule of Mixture (MROM) as there are shortcomings in cal-
culating the transverse Young’s modulus and in-plane shear mod-
ulus using SROM. Dong and Feng [12] simulate the behavior of
various unit cells using an asymptotic homogenization technique.
Other researchers have considered the uncertainty in the basic
constituent’s properties with the homogenization approach.
Kamiński and Kleiber [13] calculated the first two probabilistic
moments of the elasticity tensor using the homogenization
method of composite structure. However, due to the complicated
mathematical formulation, these studies are of limited
applicability.
The present study aims at bringing together high fidelity
micro-model based stiffness calculations of composite materials
and probabilistic analysis. The homogenization approach coupled
with a Monte Carlo Simulation method is used to predict variabil-
ity in composite material properties. The paper is organized as fol-
lows. Section 2 introduces the concept of the Representative Unit
Cell (RUC) models and the stiffness properties calculation proce-
dure using Rule of Mixture and homogenization based approaches.
Section 3 provides details of the Monte Carlo Simulation along with
Latin Hypercube Sampling and sensitivity analysis. Section 4 pre-
sents the results. Section 5 details the applicability of the method
to a wind turbine blade section analysis. Section 6 summarizes
the most important conclusions drawn from this study and high-
lights the topics require investigation in the future.
2. Theory
2.1. Representative Unit Cell (RUC) models
The wind turbine blade structure is made up of polymer-based
composite laminates, which are in turn made up of plies stacked in
a certain sequence. These plies are made up of fiber and matrix
constituents. All these levels are divided into two main groups,
the macro and micro levels as shown in the Fig. 1. Customarily,
one moves right or left in these levels via localization and homog-
enization. A homogenization procedure provides the response of a
structure given the properties of the structure’s constituents.
Conversely, the localization method provides the response of the
constituents given the response of the structure.
Macro-Level
Structure
Homogenization
Laminate
Ply
Localization
Fig. 1. Macro and micro levels.
The fibers are randomly arranged in the real unidirectional (UD)
ply. The far left side of Fig. 2 shows a cross section of a continuous
UD ply [14]. There is no obvious regular pattern in which the fibers
are arranged. A true representation of the fiber arrangement is
shown in the middle of Fig. 2. To aid computation, an idealized
fiber arrangement is used, as shown in the far right side of the
Fig. 2. In this study, two idealized RUC models are used: square
(SQR) RUC and hexagonal (HEX) RUC, as shown at the bottom of
Fig. 2. Other choices for the RUC, such as triangular RUC, could
be exploring in the future.
2.2. Rule of Mixture (ROM)
Structures made up from composite materials can be designed
by tailoring the constituent properties. This requires high fidelity
analysis and design of composite materials. Micromechanics is an
approach that handles this scenario by establishing a relationship
between the constituents and the ply or lamina. Several theoretical
models have been proposed for the prediction of composite prop-
erties from those of the constituent fiber and matrix. An investiga-
tion of the existing micromechanics models has been summarized
by Hashin [15]. The Rule of Mixtures (ROM) is one of the oldest and
simplest forms of micromechanics for calculating the mechanical
properties of unidirectional plies [7,9–10]. There is not an accurate
prediction for the transverse Young’s modulus E22 and in-plane
shear modulus G12 through SROM. This deficiency is overcome by
the Modified Rule of Mixture (MROM) as proposed by Halphin [11].
2.3. Ply stiffness computational procedure with RUC
It is not easy to determine experimentally the longitudinal and
transverse shear moduli of unidirectional composites. It is also dif-
ficult to predict these moduli using MROM as they require ply level
information [7]. Thus, numerical techniques such as the finite ele-
ment method (FEM) are needed to facilitate these predictions. The
stiffness properties of a unidirectional (UD) ply are calculated
using FEM from the constituent’s properties: the fiber and resin
(or matrix) properties; fiber volume fraction V f which controls
the geometric parameter of the RUC. For example, in the case of
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
SQR RUC, the radius of the unit cell is given as r SQR
f ¼ V f xy=p,
where x and y is lengths of square unit cell as shown in Fig. 2.
The behavior (stress and strain fields) of fiber and resin under
uniform loading is quite different due to their different material
properties. Therefore, special attention must be paid in order to
Micro-Level
Representative Unit Cell
Periodic Array
(RUC)
Matrix
Fibre
 G. Mustafa et al. / Composite Structures 131 (2015) 905–916 907

Cross Section of
Continuous Real Representation Idealized Representation
Unidirectional Ply

Glass/Epoxy (200X)*
Square RUC Hexagonal RUC Triangular RUC

y y

x z x z

Fig. 2. Representative Unit Cell models.

make sure that linear stress–strain relationships can be used to in direction in order to have stress continuity across the bound-
compute the elastic properties of the lamina. In the stiffness prop- aries of unit cell, and (3) there is no separation or overlap between
erties prediction procedure, it is assumed that in an undamaged the neighboring RUC. The displacement field boundary condition
state both constituents are perfectly bonded everywhere along on the boundary C of domain X of the unit cell is given in Eq. (2)
the length of the fiber/resin interface. For a fully reversible linear as stated by Suquet [19]:
material domain, the constitutive relationship between stress
ui ðx1 ; x2 ; x3 Þ ¼ eik xk þ ui ðx1 ; x2 ; x3 Þ ð2Þ
and strain for a UD ply is given in Eq. (1). For the plane stress case,
the stiffness matrix C is invertible to obtain a compliance matrix S. In the above, eik is the global average strain of the periodic
1 structure and xk represents a linear distributed displacement field.
Their product must produce a unity matrix, i.e. ½S ¼ ½C . By defi-
ui is a periodic part of the displacement from one RUC to another
nition, ply stiffness properties can be computed from the elements
on the boundary surface and, unfortunately, it cannot be directly
of the ½S matrix [14].
applied to the boundaries since it is unknown.
0 1 2 0 1 3
r1 C 11 C 12 C 13 0 0 0 e1 In order to determine the stiffness matrix ½C of the constitutive
Br 6
C 6C 7B Eq. (1), different displacement boundary conditions are applied on

B 2 C 6 21 C 22 C 23 0 0 0 77B
e2 C
C
B C 6 7B C the RUC with appropriate periodicity as determined by Eq. (2). The
Br  3 C 6 C 31 C 32 C 33 0 0 0 7B e3 C
B C¼6 7B C ð1Þ graphical explanation of these BCs is given in Fig. 3.
Br C B e C
B 4 C 6 6 0 0 0 C 44 0 0 77B 4C The constraint equations have been applied in the FEM code
B C 6 B C
@r 5 A 4 0 0 0 0 C 55 0 75@ e5 A ANSYS. To apply constraint equations on the nodes of opposite
r 6 0 0 0 0 0 C 66 e6 faces of the RUC, identical mesh schemes were applied on opposite
surfaces of the RUC. The degree of freedom (DOF) coupling tech-
nique in ANSYS is utilized to apply periodic boundary conditions.
2.4. Boundary conditions on RUC The coupling and constraint equations relates the motion of one
node to another [20]. For example, in the case of e13 –0, the con-
The homogenization approach acts as a bridge between the strained equations applied are shown in Fig. 4. The Von-Mises
microscopic and macroscopic scale analyses. Homogenization con- stress distribution on the RUC under BCs is also given in Fig. 4.
sists of two steps: (1) calculate local stresses and strains in con- Two important points should be noted from the RUC stress distri-
stituents (2) use homogenization to obtain global stresses and bution. First, stresses at the same location on opposite sides are the
strains for elastic property calculations. The successful implemen- same and this confirms the traction continuity. The second is that
tation of homogenization assumes that the RUC has global repeti- the boundary faces are no longer planes.
tion or periodicity. There are varieties of homogenization
approaches to predict the composite material behavior [9,16].
The homogenization technique given by Sun and Vaidya [16] is 3. Probabilistic analysis
most widely used because of its relatively low computation cost
and this can be achieved by applying proper boundary conditions 3.1. Monte Carlo Simulation (MCS) approach and Latin Hypercube
(BCs) that are periodic. These periodic BCs give more practical Sampling (LHS)
results compared with other boundary conditions such as uniform
stress boundary condition or kinematic uniform boundary condi- Most probabilistic techniques require running deterministic
tions, and have been validated by various researchers [17–18]. models multiple times with different realizations of random input
The periodic boundary conditions satisfy three statements: (1) variables. The key difference between various probabilistic tech-
the deformation should be the same on the opposite surface, (2) niques is the choice of the realization from the current analysis
the stress vectors acting on opposite surfaces should be opposite to the previous one. A probabilistic analysis answers a number of
908 G. Mustafa et al. / Composite Structures 131 (2015) 905–916

ε11 ≠ 0 ε 22 ≠ 0 ε 33 ≠ 0 ε12 ≠ 0 ε13 ≠ 0 ε 23 ≠ 0

3,Y

2,X

1,Z

Fig. 3. Boundary conditions for calculation of effective material properties of UD.

ε13 ≠ 0 l m

k n

u1x+ − u1x− = 0 u x2+ − u x2− = 0 u x3+ − u x3− = 0

u1y+ − u1y− = 0 u 2y + − u 2y − = ε12d nk u 3y+ − u 3y− = 0
u1z+ − u1z− = 0 uz2+ − uz2− = 0 uz3+ − uz3− = ε12d mn
Fig. 4. Shear boundary conditions on RUC.

questions: (1) how much scatter induced due to randomness in the
input variables; (2) what is the probability the output parameters
are no longer fulfilled based on design criterion; (3) which random
input variable most affect the output parameters which helps to
screen the design variables [20]?. The Monte Carlo Simulation
(MCS) approach is one of the most general tools to perform
stochastic analysis under uncertainty of input variables [21]. In this
work, the second order statistics of the responses are obtained
using MCS with the random input variables (RVs) specified by their
mean l and standard deviation r. The MCS consists of three steps:
(1) generation of realization corresponding to probability distribu-
tion function (PDF) – Latin Hypercube Sampling is used for this
purpose in this work; (2) a finite element simulation evaluates
responses for each realization; (3) statistical analysis of the results
yields valuable information concerning the sensitivities of the
responses to the stochastic RVs. In the Latin Hypercube Sampling
technique the range of all random input variables is divided into
n intervals with equal probability and the sample generation pro-
cess has a memory in the meaning that the sampling points cannot
cluster together.
3.2. Sensitivity analysis
Analysis of parameter sensitivity investigates the effect of vari-
ability of certain input random variables on the variability of
design-relevant response quantities; the sensitivities can be
described by coefficients of correlation. The most commonly used
coefficient of correlation proposed by Bravais-Pearson is given in
Eq. (3) between two random variables X and Y:
Cov ðX; YÞ
r X;Y ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffipffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
VarðXÞ VarðYÞ
The flow chart to calculate probabilistic properties of a UD com-
posite based on the constituent’s uncertainties in a modular fash-
ion is illustrated in Fig. 5.
4. Results and discussion
4.1. Boundary conditions for 2D unit cell
The periodic boundary conditions (BCs) devised by Xia et al.
[18] and given in Eq. (4) are applied to a SQR and HEX unit cell
FEM model which insures that the composite has the same defor-
mation mode and there is no separation between unit cells. To ver-
ify the BCs, a 2-dimensional unit cell model is considered first. The
fiber volume fraction is 50%. The elastic moduli and Poisson’s ratio
f
for the fiber and matrix are E11 ¼ 72500 MPa; v 12f ¼ 0:22 and
m
E ¼ 2600 MPa; v ¼ 0:42, respectively.
m
ujþ j j
i ðx; y; zÞ  ui ðx; y; zÞ ¼ c i ; ði; j ¼ x; y; zÞ ð4Þ
In Eq. (4), ujþ j
i and ui are displacements of one pair of nodes at
opposite boundary faces, with identical coordinates in the other
two directions. The constant cij represent the stretch or contraction
of the unit cell model under the action of normal forces or shear
deformations due to the shear forces. A simple fiber–matrix
domain was defined in ANSYS [20] tool as shown in Fig. 6. The
FEM unit cell model and the deformed shape of the unit cell are
shown in Fig. 6. The boundaries do not remain planes after load
application, and shear stresses are uniform in the unit cell.
Table 1 compares the present results with Xia et al. [18] demon-
strating excellent agreement of shear strain, shear stress, and
ð3Þ equivalent shear modulus.
 G. Mustafa et al. / Composite Structures 131 (2015) 905–916 909

User Subroutine O/P Parameters

I/P Random Variables MCS RUC + PBCs EUD(μ, σ)
E f (μ, σ), E m (μ, σ) Latin Hypercube Sampling u x2+ − u x2− = 0
u − u 2y − = ε12d nk
2+
y

No Convergence
μ, σ
Yes

Post Processing

Statistical Response of Correlations

EUD Coefficients

Fig. 5. Probabilistic analysis flow chart.

2D Unit Cell Shear Stress τxy (MPa) - Plot

E A E A
Matrix
Fiber

Fiber

Y (v)
F
B
F B
X (u)

Fig. 6. 2D Unit cell and shear stress distribution.

Table 1 Table 2
2D Unit cell model results. Material properties of fiber and matrix.

Parameter Xia et al. [18] FEM Properties E-glass fiber Properties MY750 Epoxy

Shear stress, sxy (MPa) 6.4831 6.4917

f
E11 (GPa) 74.0 Em (GPa) 3.35
Shear strain, cxy (–) 0.0036 0.0036 f
E22 f
, E33 (GPa) 74.0 vm (–) 0.35
Shear modulus, G (MPa) 1801 1803 f f 30.8
G12 , G13 (GPa)
vf 0.2
12 (–)

4.2. Boundary conditions for 3D unit cell

The next step was to define a finite element model of a square
unit cell and repeated array in order to compare the behavior
under identical loadings and boundary conditions. The 3-D struc-
tural solid element, SOLID45, is used for the FEM analysis. The
material properties of Silenka E-Glass 1200 tex and MY750 Epoxy
are given in Table 2.
The responses of the multi-cell array and the unit cell model
were compared for different loading conditions with proper peri-
odic boundary conditions. As an example, the comparison of the
SQR and HEX multi-cell array and the unit cell models under shear
load is shown in Fig. 7. From Fig. 7, it is clear that the stress distri-
butions in the multi-cell model and the unit cell are identical.
There is no boundary separation, which verifies the boundary con-
ditions. The rest of the boundary conditions were verified in a sim-
ilar manner.
4.3. UD ply stiffness properties
To verify the homogenization technique, the stiffness properties
were determined with the SQR and HEX unit cell first and com-
pared with test data [22] as well as predictions from SROM and
MROM. The same E-glass fiber and MY750 epoxy was used for this
analysis. The stiffness properties were calculated in the linear elas-
tic region. The comparisons were made using 60% fiber volume
910 G. Mustafa et al. / Composite Structures 131 (2015) 905–916

SQR

0.3106 1.4245 0.3104 1.4241

HEX

0.3491 1.5468 0.3489 1.5465

Fig. 7. Comparison of shear stress (N/m2) distribution in multi-cell & unit cell model.

60
Test Data SQR
50
UD Ply Sffness (GPa)

HEX HEX SROM

MROM
40
SQR
30

20

10

0
E11 E22 E33

8
0.50 Test Data SQR
UD Ply Shear Sffness (GPa)

Test Data SQR 7 HEX SROM

UD Ply Poisson's Rao (-)

HEX SROM 6 MROM

0.40
MROM
5
0.30
4
0.20 3
2
0.10
1
0.00 0
v12 v13 v23 G12 G13 G23

Fig. 8. Comparison of stiffness properties.

fraction. The predictions were compared with test data [22] and
were found to be in good agreement, as shown in Fig. 8.
4.4. Statistical model description and Monte Carlo Simulation
In this section, the Monte Carlo Simulation approach is coupled
with a homogenization based stiffness calculation procedure to
access the uncertainty of the stiffness properties of UD due to vari-
ation in the constituent’s properties. This coupling is achieved in
three steps: (1) generation of realizations according to the Latin
Hypercube Sampling technique of random variables, i.e. the fiber
and the matrix; (2) using the RUC and homogenization approach
to simulate UD properties against each set of realizations; (3)
post-processing the results to calculate statistical response and
sensitivities of the output response parameters. Steps 1 and 2 were
implemented in the finite element tool ANSYS PDS and ANSYS
Multiphysics respectively by user defined subroutines. Step 3 is
performed partially in ANSYS PDS and in MATLAB. The generation
of the realizations is based on the probability distribution; a
Gaussian distribution is used for fiber and matrix property random
variables. Note that an alternate choice of the distribution type
would not modify the overall methodology. The Gaussian probabil-
ity density function (PDF) is given in Eq. (5). The mean lX and stan-
dard deviation rX are the two parameters describing the Gaussian
PDF for each random variable X.
" #
1 ðx  lX Þ2
f X ¼ pffiffiffiffiffiffiffiffiffiffiffiffi exp  ð5Þ
2prX 2r2X
 G. Mustafa et al. / Composite Structures 131 (2015) 905–916 911

Table 3 and output response parameters. The results of this analysis are
Input random variables for calculation of UD properties. shown in the Fig. 10. These correlations explain the strength of
Material Random variable Symbols Mean Std dev the relationship between the stochastic quantities and range from
Fiber Longitudinal modulus (GPa) f 74.00 18.50 1 to +1. A correlation of 1 describes perfect negative relation
E11
Transverse modulus (GPa) f 74.00 14.80
and +1 a strong positive relation. A correlation close to 0 indicates
E22
Transverse modulus (GPa) f 74.00 14.80
no or weak relation. It can be clearly seen from Fig. 10 that UD lon-
E33
gitudinal stiffness E11 has a strong relation with fiber longitudinal
In-plane shear modulus (GPa) f
G12 30.80 7.70
modulus. Likewise, E22 has a strong positive relation (close to 1)
Transverse shear modulus (GPa) f
G13 30.80 6.16
with matrix modulus Em . The shear properties of composites are
Transverse shear modulus (GPa) f
G23 30.80 6.16
also very impotent, particularly in wind turbine blade shear webs
Major Poisson’s ratio (–) v f 0.20 0.01
12 [24]. The sensitivity analysis shows that matrix shear modulus
Minor Poisson’s ratio (–) v f 0.23 0.01
13 Gm correlates very highly with UD lamina shear molulii, i.e.
Minor Poisson’s ratio (–) v f 0.23 0.01
23 G12 ; G13 ; G23 but Gm has only a minor sensitivity to the UD lamina
m
Matrix Elastic Modulus (GPa) E 3.35 0.84 Poisson’s ratio. This sensitivity analysis aids in improving the
Poisson’s ratio (–) vm 0.35 0.02 material according to design requirements as well as screening
Shear modulus (GPa) Gm 1.24 0.25
out unnecessary design variables for optimization purposes.
The statistical interdependence between the output response
parameters was also calculated by a Spearman Rank Order correla-
The details of the constituent random variables are given in the tion coefficients approach. Fig. 11 shows the sensitivity plot of the
Table 3. The properties of the constituents (E-glass fiber and response parameters, i.e. UD stiffness properties. It can be seen
MY750 matrix) were taken from [22]. A 5% standard deviation is from this Fig. 11, some properties are strongly correlated with
considered where data was missing for a particular quantity. The others and some are in a week manner. This is important informa-
response parameters include: longitudinal modulus (E11 ), trans- tion for further analysis of the wind turbine blade. At the time of
verse modulus (E22 ) and (E33 ), in-plane shear modulus (G12 ) and declaring the stochastic properties of UD for the analysis of blade,
(G13 ), transverse shear modulus (G23 ), major Poisson’s ratio (v 12 ), a threshold criterion will be applied between correlations in UD
and minor Poisson’s ratio (v 13 ) and (v 23 ). properties. This will not only accelerate the overall simulation time
A common question is the required number of iterations to be but also give more insight into structural response dependency on
performed with the Monte Carlo method. Eq. (6) [23] can be used the input variables as explained in the following section.
to estimate the required number of runs by the standard error of
the mean of the distribution relation:
5. Application to wind turbine blade section
zrX
error ¼ pffiffiffi ð6Þ
n A case study was carried out to highlight the implications of
micromechanics-based probabilistic analysis for realistic applica-
where z is a confidence multiplier and its value is 2 for 95% confi-
tion to a composite wind turbine blade structure. Wind turbine
dence level, rX is standard deviation and n is the is the number of
blades can be analyzed as beam-like structures, as is done by cur-
runs or simulations performed in the Monte Carlo Simulation. The
rent design codes [6,25]. In this section though, the analysed blade
percentage error of the mean becomes:
is modeled in more detail as a shell structure. The investigated
100zrX blade has shell and spar/web type internal structural layout. The
E¼ pffiffiffi ð7Þ
lX n aerodynamic outer shell is obvious as it is providing required aero-
dynamic characteristics to the blade. The internal blade structure
For the estimation of runs n, the Eq. (7) becomes:
consists of spar caps and shear webs. The spar caps carry the bend-
 2
100zrX ing and axial loads. The shear webs carry the shear stress and pro-
n¼ ð8Þ vide airfoil shape stability. The classic embodiment of this concept
lX E
is similar to the steel I-beam except that there are shells around
For example, using UD longitudinal stiffness (E11 ) from test the outside that form the aerodynamic shape. The geometry and
data, lE11 ¼ 45:6 GPa and rE11 ¼ 9:12 GPa, for a 95% confidence other specifications of the rotor blade considered in this research
level z ¼ 2 and E ¼ 2, the required number of Monte Carlo is based on the NREL 5 MW baseline wind turbine described in
Simulations is 400. [26]. A 61 m blade is attached to a hub with a radius of 2 m, which
The results of the probabilistic analysis after post-processing in gives the total rotor radius of 63 m. Although the blade is com-
MATLAB are given in Fig. 9. The response parameters are presented posed of several airfoil types, in this work the airfoil used is the
as histograms along with the PDFs fit to the data; Gaussian was NREL-S818 [27] from maximum chord to the tip of the blade.
found to be a good fit for the UD properties. The higher standard The first portion of the wind blade is a cylindrical in shape.
deviation in the stiffness property of the fiber is reflected in the Further away from the root the cylinder is smoothly blended into
UD longitudinal property response parameter. The probability that a NREL-S818 airfoil at maximum chord. The power curve and other
the UD longitudinal stiffness E11 is smaller than 35 GPa is 0.16677 environmental parameters are given in Fig. 12.
with a 95% confidence level. The lower stiffness of E-glass/epoxy A part of the blade is considered for simplicity of computation.
composites will cause more deflection in the blade. Therefore, it The details of the blade wing box and the composite layup
is important to calculate this probability. The standard deviation sequence are given in the Fig. 13. The investigated blade consists
in UD transverse stiffness is 3.00 GPa. The standard deviations in of three different types of E-glass/epoxy composites: unidirectional
shear modulii and Poisson’s ratio are given in the Fig. 9. (UD) plies, BX [±45°]s, and TX [0°/±45°]s. The UD and TX are used in
The evaluation of the probabilistic sensitivities is based on the the shell structure along with a PVC core. The BX and PVC core [28]
correlation coefficients between all random input variables and a is used in shear webs to provide shear resistance. The properties of
particular random output parameter. The sensitivity analysis was the materials used in the wing box analysis were calculated from
performed using the Spearman rank coefficient of correlation the probabilistic analysis on RUC as described in the previous
approach to investigate correlation between input design variables section.
912 G. Mustafa et al. / Composite Structures 131 (2015) 905–916

Fig. 9. Histograms & PDFs of output parameters.

parked wind turbine is either in standstill or idling condition. For

an annual average wind of 12 m/s, the 50-year extreme wind is
70 m/s which is the maximum of the 3s average. The left side of
Fig. 14 shows the FEM results under one set of input realizations
and the maximum deflection calculated with this set of realiza-
tions was 195.1 mm.
Fig. 15 illustrates the procedure used to calculate the proba-
bilistic response of the wing box caused by uncertainties in the

Fig. 10. Spearman Rank Order correlation coefficients.

The finite element model of the wing box is shown in the

Fig. 14. This model was generated in the ANSYS Multiphysics mod-
ule. The wing box was modeled using the SHELL181 ANSYS ele-
ment type [20]. Two types of loads were applied: flapwise and
edgewise loads. These loads were calculated using an aeroelastic
computer-aided engineering tool for horizontal axis wind turbines,
FAST [29]. Fig. 14 shows the boundary conditions applied to the
wing box section. The quasi-static analysis were performed accord- Fig. 11. Spearman Rank Order correlation coefficient matrix between output
ing to IEC-DLC 6.1 load case [6]. In this load scenario, the rotor of a parameters.
 G. Mustafa et al. / Composite Structures 131 (2015) 905–916 913

Rated Power = 5 MW Parameter Units NREL-5M

D = 126 m Rated Power MW 5
Vrated = 12 m/s IEC Class (-) 1A
Number of blades (-) 3
Wind Ω= 13rpm Vin = 3 m/s Rated wind speed m/s 12
Vout = 25 m/s
Cut-in wind speed m/s 3
Cut-out wind speed m/s 25
Z0 = Surface roughness parameter Hub height = 110 m Blade Tip Speed m/s 81.7
Vin = Cut-in wind speed
Vout = Cut-out wind speed Rated Rotor Speed rpm 13
Ω = Angular speed of rotor
Tip Speed Ratio (-) 6.81
Diameter m 126
6000
Power Curve Blade Length m 63
Power Output (kW)

5000

Cut-out Wind
Hub Height m 110

Speed
4000
Maximum Chord m 4
3000 Cut-in Wind
Speed Maximum Twist deg 14
2000
Precone deg 2.5
1000
Rated Power = 5.0 MW
Direction (-) Clockwise
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Orientattion (-) Upwind
Wind Speed (m/s)

Fig. 12. NREL-5 MW baseline wind turbine.

Spar caps

Lay-up Sequence Lay-up Sequence

@ Spar Caps @ Shear Webs
1.5 m
TX [0/ 45]S
2 Plies TX 5 Plies BX BX [ 45]S
30 Plies UD 8 mm PVC
2 Plies TX 5 Plies BX

Shear Webs

Fig. 13. Blade wing box and stacking sequence.

UD properties in a modular manner. These modules consist of: (1)
selection of input design variables and description of their mean
and standard deviation along with distribution type; (2) realization
generation using efficient Latin Hypercube Sampling (LHS)
approach; (3) user subroutine is used to calculate stiffness proper-
ties of UD with micro model and periodic boundary conditions
(PBCs); (4) post processing of data performed once convergence
criteria on mean values and standard deviations is met; (5) filters
are applied with a threshold of 35% on the correlations given in
Fig. 11 to screen the unimportant design variables; (6) probabilistic
finite element analysis were performed on the wing box with fil-
tered correlation matrix and stochastic properties of UD. The
results of step 6 were then post-processed to calculate the overall
probabilistic response of the wing box.
A summary of the stochastic response of the wing box deflec-
tion is shown in the Fig. 16 with and without considering correla-
tion and the fitted PDF on the right side of the Fig. 16. The PDF fit to
the data is a Weibull distribution, even though the distributions of
inputs were Gaussian. The 2-parameter Weibull model used to fit
the data.
 x b1
b
fX ¼ exp eðx=aÞb ð9Þ
a a
One of the Weibull model parameter a controls the scale and the
second one b controls shape. Several methods have been devised
to estimate the parameters that will fit the particular data distribu-
tion. Some available parameter estimation methods include proba-
bility plotting, rank regression, and maximum likelihood estimation
914 G. Mustafa et al. / Composite Structures 131 (2015) 905–916

Fig. 14. Boundary conditions & FEM modeling of wing box.

Fig. 15. Flow chart for probabilistic analysis of wing box.

60 0.014
Without Correlations Without Correlations Weibull Distribution Without Correlations
: 200.33 With Correlations
With Correlations
Distribution Overlap : 27.35 Without Correlations
: 4.07 Weibull Parameters
0.012 : 212.61
50 : 6.92
With Correlations
: 200.45
: 27.70
Probability Density Function

With Correlations
: 4.03 0.01 Weibull Parameters
: 212.87
40 : 6.85
No. of Occurence

0.008

30

0.006

20
0.004

10
0.002

0 0
140 160 180 200 220 240 260 280 300 320 140 160 180 200 220 240 260 280 300 320
Wing Box Displacement at Free End (mm) Wing Box Displacement at Free End (mm)

Fig. 16. Wing box deflection with and without correlation.

 G. Mustafa et al. / Composite Structures 131 (2015) 905–916
With Correlation
0 0
Significant:
E1_EQ
-.1 G12_EQ
G13_EQ
G23_EQ
Sensitivity Level
NU23_EQ
Sensitivity Level
-.2
E3_EQ
E2_EQ
NU12_EQ
-.3 NU13_EQ
Significance level:
-.4 2.500%
95%
-.5
-.6
-.7
-.8
Output Parameter MAX_DISP
Fig. 17. Wing box deflection sensitivities with and without correlation.
(MLE) [30–31]. The MLE approach in MATLAB was used to estimate
the Weibull parameters. The Weibull parameters are given in the
Fig. 16. The MCS applied in this work used 500 realizations for prob-
abilistic analysis of the blade section, without the use of a MCS con-
vergence criterion. In order to have a fair comparison between
results with and without including correlations, the same seed
number was used to generate the random realizations. As can be
seen from the histogram on the left side and the fit PDF on the right
side of Fig. 16, the correlations have little effect on the maximum
displacement of the wing box. However, inclusion of the correlation
coefficients highlighted in detail the important factors on the
deflection of wing box, as shown in Fig. 17 which shows the sensi-
tivities on output response parameter with a 95% significance level.
Based on these results, ignoring correlations suppressed the sensi-
tivities of input variables on the response parameters. The consider-
ation of correlations helps to screen some of the unimportant
factors such as E22 ; E33 ; v 12 ; E13 with a threshold of 20%. These
parameters can therefore be treated as deterministic in the future
reliability analysis. Also, from the sensitivity analysis, one can con-
clude that E11 and G12 are the two key input variables that controls
the deflection and these variables are related to the output deflec-
tion response in a strong negative manner, which means with the
increase of these variables, the deflection response will be lower.
Overall, these analyses illustrate the capability of the coupled
approach for structural robust analysis by indicating the variables
to change in order to achieve a given reliability level.
6. Conclusion
This work aimed to develop a coupled approach for stiffness
property prediction of composite materials used in wind turbine
blades using advanced micromechanics and reliability-based
methodologies. The homogenization approach was used with a
unit cell to estimate composite material stiffness properties. The
predicted results were compared with SROM and MROM along
with the test data and found to be in good agreement with test
data and MROM. The work was then extended to perform the prob-
abilistic analysis with Monte Carlo Simulation to incorporate
uncertainties in the constituent’s properties. Latin Hypercube
Sampling was employed to cover most of the input variable design
space. From the probabilistic analysis, it was found that the equiv-
alent properties of UD followed the Gaussian distribution and
these properties were affected by variations in fiber and matrix
properties. The Spearman Rank Order sensitivity analysis gave
insights into important constituent’s properties and found that
the modulus of fiber and matrix has more influence on composite
915
Without Correlation
Significant:
E1_EQ
-.1 G12_EQ
Insignificant:
E2_EQ
-.2
E3_EQ
G13_EQ
G23_EQ
-.3 NU12_EQ
NU13_EQ
NU23_EQ
-.4
Significance level:
2.500%
-.5 95%
-.6
-.7
-.8
Output Parameter MAX_DISP
properties. In addition, correlations were calculated between UD
properties which were then used in blade analysis. This probabilis-
tic micromechanical approach for composite materials provides a
useful tool for performing preliminary material design for finaliz-
ing material for structural application.
From a practical point of view, the method was then applied to
a 5 MW wind turbine blade structural analysis. A section of the
blade was considered for demonstration purposes. The stochastic
finite element analysis was performed in ANSYS/Multiphysics
using a user subroutine in order to simulate probabilistic response
of wing box section. Two analyses were simulated with and with-
out correlations between UD properties. There was no significant
difference between PDFs but ignoring correlations suppressed the
sensitivities of input variables on the response of wing box. Also,
it was found that longitudinal modulus and in-plane shear modu-
lus of the composite are the most critical material properties that
influence deflection of the blade box. Besides this, sensitivity anal-
ysis screens out unimportant input variables which can be treated
as deterministic in the further analysis. Furthermore, sensitivity
analysis highlights the key material properties that most influence
the response of the structure and this will eliminate/reduce unnec-
essary time consuming and expensive full testing campaign. In
future work, the proposed probabilistic analysis can help deter-
mine constraints for use in structural optimization. The next step
in this research is to expand micromechanics based probabilistic
analysis to estimate strength and fatigue life of composites using
Bayesian Inference approach.
Acknowledgements
The authors would like to thank the Natural Sciences and
Engineering Research Council of Canada (NSERC) and Composite
Research Network (CRN) for financial support of this work.
References
[1] Veers PS, Ashwill T. Trends in the design, manufacture and evaluation of wind
turbine blades: Library WO; 2003.
[2] Henderson AR, Morgan C. Offshore wind energy in europe— A review of the
state-of-the-art; 2003.
[3] Hillmer B, Borstelmann T. Aerodynamic and structural design of multiMW
wind turbine blades beyond 5MW; 2007.
[4] Fingersh LJ, Hand MM, Laxson AS. Wind turbine design cost and scaling model:
National Renewable Energy Laboratory Golden, CO; 2006.
[5] Kam T, Tsai S, Chu K. Fatigue reliability analysis of composite laminates under
spectrum stress. Int J Solids Struct 1997;34:1441–61.
[6] International Electrotechnical Commission. IEC 61400–1 3rd ed. 2005–08
Wind turbines – Part 1: Design requirements; 2005.
[7] Tsai SW. Theory of composites design: Think composites Dayton; 1992.
916 G. Mustafa et al. / Composite Structures 131 (2015) 905–916
[8] Tsai SW. Strength & life of composites: Stanford University; 2008.
[9] Hashin Z, Rosen BW. The elastic moduli of fiber-reinforced materials. J Appl
Mech 1964;31:223–32.
[10] Liu G. A step-by-step method of rule-of-mixture of fiber-and particle-
reinforced composite materials. Compos Struct 1997;40:313–22.
[11] Halpin JC. Effects of environmental factors on composite materials; 1969.
[12] J-w Dong, M-l Feng. Asymptotic expansion homogenization for simulating
progressive damage of 3D braided composites. Compos Struct
2010;92:873–82.
[13] Kamiński M, Kleiber M. Perturbation based stochastic finite element method
for homogenization of two-phase elastic composites. Comput Struct
2000;78:811–26.
[14] Tsai SW. Theory of composite design dayton. Ohio; 1992.
[15] Hashin Z. Analysis of composite materials—a survey. J Appl Mech
1983;50:481–505.
[16] Sun C, Vaidya R. Prediction of composite properties from a representative
volume element. Compos Sci Technol 1996;56:171–9.
[17] Kouznetsova V, Brekelmans W, Baaijens F. An approach to micro-macro
modeling of heterogeneous materials. Comput Mech 2001;27:37–48.
[18] Xia Z, Zhang Y, Ellyin F. A unified periodical boundary conditions for
representative volume elements of composites and applications. Int J Solids
Struct. Elsevier; 2003. Pp. 1907-21.
[19] Suquet P. Elements of homogenization theory for inelastic solid mechanics.
Homogenization Tech Compos Media 1987:1994. 1994-275.
[20] ANSYS. Ansys Multiphysics Release 14.0. 2013.
[21] Ayyub BM, Haldar A. Practical structural reliability techniques. J Struct Eng
1984;110:1707–24.
[22] Soden P, Hinton M, Kaddour A. Lamina properties, lay-up configurations and
loading conditions for a range of fibre-reinforced composite laminates.
Compos Sci Technol. Elsevier; 1998. Pp. 1011-22.
[23] Driels MR, Shin YS. Determining the number of iterations for Monte Carlo
simulations of weapon effectiveness; 2004.
[24] Burton T, Jenkins N, Sharpe D, Bossanyi E. Wind energy handbook. John Wiley
& Sons; 2011.
[25] Lloyd G. Rules and regulations, IV-Non-marine technology. Part 1-Wind
Energy; 1993.
[26] Jonkman JM, Butterfield S, Musial W, Scott G. Definition of a 5-MW reference
wind turbine for offshore system development: National Renewable Energy
Laboratory Golden, COL 2009.
[27] Somers D. The S816, S817, and S818 Airfoils. NREL Contractor Report NREL/SR-
500-36333; 2004.
[28] Gurit. Gurit PVC Structural Foam Core; 2014.
[29] Jonkman JM, Buhl M. FAST User’s Guide. National Renewable Energy
Laboratory; 2005.
[30] Engelhardt M. On simple estimation of the parameters of the Weibull or
extreme-value distribution. Technometrics 1975;17:369–74.
[31] Al-Fawzan MA. Methods for estimating the parameters of the Weibull
distribution. King Abdulaziz city for science and technology Saudi Arabia;
2000.